Don’t Teach the Robots to Lie

By Greg Lukianoff & Adam Goldstein

Monday, July 21, 2025

 

Texas recently became the second state to pass a law banning algorithmic discrimination — that is, sanctioning operators of AI systems if those systems unfairly, for example, deny you a bank loan, a second job interview, or an insurance policy. I hope there is no third. These regulations could be catastrophic for truth, innovation, and, ultimately, human freedom.

 

The Texas Responsible AI Governance Act, which was signed into law last month and will be effective next January, prohibits intentionally creating or using AI in a way that would “infringe, restrict, or impair” a constitutional right or discriminate on the basis of a protected class. Anyone found to have created such an AI faces potential fines and enforcement from the state attorney general. Requiring the element of intent at least sets it slightly apart from Colorado’s law, which was passed last year and will be effective next February.

 

Under Colorado’s Consumer Protections for Artificial Intelligence law, any “high-risk” AI system that plays a “substantial role” in a “consequential decision” must undergo rigorous annual and ongoing impact assessments, risk mitigation planning, and continual audits for potential bias. A “consequential decision” is one that has a “material legal or similarly significant effect” on the provision or denial of services. In other words, one that actually affects the outcome. And if the AI system fails — or there’s an oversight — it could mean a lawsuit, a regulatory fine, or worse. In the words of Governor Jared Polis, the law regulates “the results of AI system use, regardless of intent.” Compliance with the law creates a rebuttable presumption of reasonable care.

 

Both laws are enforced exclusively by their respective state attorneys general — in other words, they don’t allow individuals to sue over alleged violations. In Texas, the attorney general’s office is required to create a public portal to accept complaints; penalties could be as high as $200,000, or $40,000 a day for ongoing violations. Violations of Colorado’s law are addressed under the state’s Unfair or Deceptive Trade Practices Act, which authorizes penalties of up to $20,000 per violation.

 

In theory, these laws prohibit AI systems from engaging in algorithmic discrimination — or from treating different groups unfairly. In effect, however, AI developers and deployers will now have to anticipate every conceivable disparate impact their systems might generate and scrub their tools accordingly. That could mean that developers will have to train their models to avoid uncomfortable truths and to ensure that their every answer sounds like it was created with HR and legal counsel looking over their shoulder, softening and obfuscating outputs to avoid anything potentially hurtful or actionable. In short, we will be (expensively) teaching machines to lie to us when the truth might be upsetting.

 

This polite, legally sanitized dishonesty would be enforced, by the power of the state, on every algorithm that might influence your life. That includes the résumé filter that screens your job application, the chatbot that gives you advice, and even the underwriting system that decides your mortgage rate.

 

There is a moral panic driving this type of legislation: the idea that if a model says something unkind or produces unequal results, it must be a threat to civil rights. But systems that are optimized for litigation rather than reality should concern us even more. When you teach machines that truth is subordinate to legal compliance, you’re not making the world safer — you’re just hiding the evidence of risk under a blanket of euphemism and suppression. It would be similar to applying the maxim “Don’t discuss politics, sex, or religion in polite company” to everything we can or should care about, everywhere, all the time.

 

This is a theme we’ve been harping on for years: We’ve become a society more concerned with appearing just than with doing justice. We punish plain speaking as impolite. We reward bureaucratic cover-your-assets maneuvers as leadership. And now, with AI, we have the opportunity either to break this pattern or to lock it in for the long term.

 

Algorithmic discrimination laws bake into code a kind of cowardice that already plagues too many of our public and private institutions. At Harvard, for example — whose motto is Veritas — professor Roland Fryer was functionally exiled by his peers for his controversial research, including a study that found that black suspects were less likely to be shot by police than white suspects. Also at Harvard, professor Carole Hooven was pushed out of her job for suggesting that the terms “man” and “woman” are useful in discussions of biology. Hooven’s research on sex differences has taken on particular significance in the wake of the Supreme Court’s decision upholding state bans on so-called gender-affirming care.

 

Our organization, the Foundation for Individual Rights and Expression, has tracked how campus antidiscrimination efforts are deployed to silence dissent for over two decades. While FIRE was not around for the first great age of “political correctness” in the 1980s, since our founding in 1999, we have seen rules to stop racial harassment abused to stop reading; laws to stop gender harassment stretched to police “offensive content”; and public-health rules to justify silencing policy critics.

 

Now we are imposing this same pressure to conform onto the systems that will dominate the future of our search for truth. As a result, the marketplace of ideas in tech will start to look like the one in higher education. And the further AI drifts from telling users the truth, the less useful it will be and the less likely it is that these solutions will be deployed.

 

That’s not good for anyone.

 

Yes, AI systems have the potential to generate algorithmic discrimination. But subjecting the use of AI systems to the ongoing threat of government punishment for “wrong ideas” is like threatening to close libraries if anybody gets the wrong message from a dog-eared copy of Animal Farm.

 

Colorado’s law doesn’t burden only Big Tech. It applies to any company that uses AI to help make decisions — from mom-and-pop landlords to local school districts. And it makes everyone in the software supply chain jointly liable for discrimination even if they are three steps removed from some decision that has been deemed discriminatory. This creates an incentive to clamp down, to self-censor, to overburden innovation with layers of red tape and litigation.

 

Unfortunately, states are lining up to pass their own laws. Connecticut’s senate passed a bill similar to the laws in Colorado and Texas, and more than 20 other states are considering doing the same, including Virginia, Utah, and California. Other rules, with narrower scopes — such as ones focused on employment discrimination — have been proposed or promulgated in Illinois, New Jersey, and New York. Congress had been considering freezing the legal landscape with a ten-year ban on state-level AI regulation, as part of its recent reconciliation negotiations; that later became five years, and then the provision was dropped entirely.

 

A patchwork of state laws could create regulatory chaos. Texas’s and Colorado’s laws purport to regulate the use of AI that affects residents of their states regardless of where the servers or company in any given case might be located. In the context of age-verification regulations, we have seen some websites decide to simply ban visitors from states with onerous laws. But individuals’ losing access to some adult websites is substantially less consequential than would be a loss of access to, say, insurance coverage or banking products.

 

Some who champion such laws seem to fear that our machines are becoming smarter than us. We don’t share that fear. What terrifies us is that, if they become smarter than us, we’ll have taught them our worst habits — our fear of speaking clearly, our obsession with avoiding blame, our willingness to twist facts into whatever shape fits the current moral fashion. We should be teaching AI how to avoid hallucinations, not curating its hallucinations to ensure that they’re politically correct.

 

Politeness is a virtue, sure. But it’s not one worth sacrificing reality for. If AI systems produce outcomes that are used by humans in discriminatory ways, we have traditional laws to stop that. But teaching machines that may soon surpass us in intelligence that it’s acceptable — or even necessary — to bend the truth in order to avoid legal exposure or hurt feelings is madness. We need AI that is honest, not AI that has been trained by 500 risk analysts and a DEI consultant to never say anything that might provoke a lawsuit.

 

A relentless commitment to truth is no longer just an academic value. It is existentially important. If we force AI companies to distort reality to satisfy vague standards of “fairness” or “disparate impact,” we won’t be preventing Skynet. We’ll be programming it.